Research
Transform Trained Transformer for Accelerating Native 4K Video Generation
arXiv:2512.13492v2 Announce Type: replace Abstract: Native 4K (2176imes3840) video generation remains a critical challenge due to the quadratic computational explosion of full-attention as spatiotempo
arXiv:2512.13492v2 Announce Type: replace Abstract: Native 4K (2176imes3840) video generation remains a critical challenge due to the quadratic computational explosion of full-attention as spatiotemporal resolution increases, making it difficult for models to strike a balance between efficiency and quality. This paper proposes a novel Transformer retrofit strategy termed T3 (extbf{T}ransform extbf{T}rained extbf{T}ransformer) that, without altering the core architecture of full-attention pretrained models, significantly reduces compute requirements by optimizing their forward logic. Specifically, extbf{T3-Video} introduces a multi-scale weight-sharing window attention mechanism and, via hierarchical blocking together with an axis-preserving full-attention design, can effect an "attention pattern" transformation of a pretrained model using only modest compute and data. Results on extbf{4K-VBench} show that extbf{T3-Video} substantially outperforms existing approaches: while delivering performance improvements (+4.29parrow VQA and +0.08parrow VTC), it accelerates native 4K video generation by more than 10imes. Project page at https://zhangzjn.github.io/projects/T3-Video
Related
- NABLA: Neighborhood Adaptive Block-Level Attention
- Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation
- Sparse VideoGen2: Accelerate Video Generation with Sparse Attention via Semantic-Aware Permutation
- Mixture of Distributions Matters: Dynamic Sparse Attention for Efficient Video Diffusion Transformers
- CineScale: Tuning-Free High-Resolution Video Generation
Source: arXiv cs.CV | 2026-10-06